DHH: The Real Bottleneck in Software Isn't AI. It's Management.

Appearing on the Lex Fridman Podcast, Basecamp co-founder David Heinemeier Hansson made a claim aimed squarely at large software companies: the reason legacy products like Photoshop or Premiere aren't shipping features faster in the AI era has nothing to do with the capability of coding agents. It has to do with corporate hierarchy.

Hansson, who has spent three months building a project he calls Amachi, said the biggest productivity gains from AI agents — 10x, 100x, in rare cases 1,000x — only materialize when a person works with the agent directly, with no human intermediary. Layer in a product manager, designers, a VP and a CTO, each wanting a say in the process, and the gains evaporate. "That's where all the productivity goes to die," he said. His broader point: in large organizations, implementation was rarely the constraint to begin with. Human bandwidth and communication were.

Hansson also pushed back on the idea that AI-generated code is inherently sloppier than human-written code. Asked whether current tools require a trained programmer to "vibe code" — a term for prompting an AI to generate working software with minimal manual coding — on existing, complex codebases, he said yes, if you want to preserve architecture. But he rejected the premise that human programmers do better by default: "have you looked at the average programmer's output?" he said, describing codebases that have passed through thousands of engineers at major companies as "awful."

His sharper claim was about vision, not code quality. Most companies, he argued, aren't bottlenecked on implementation capacity — they're bottlenecked on ideas, taste and judgment about what to build. Handing agents to an organization that lacks those things, he said, just produces more bad ideas, faster, with nothing worth shipping at the end of it. He did not name specific companies as examples of this failure mode beyond a passing jab at Microsoft's development pace.

The exchange comes as coding-agent tools proliferate and enterprises debate how much of software development can be automated. Hansson's framing puts the constraint on management structure rather than model capability — a claim from Hansson, not an established consensus.

Heard on Lex Fridman Podcast — #501 – DHH: Future of Programming, AI, Agentic Engineering, Vibe Coding & Linux (2026-08-26).